Build It and They Will Come: How Hardware Influences Large-Scale Simulation Science at Extreme Scales: Where Big Data Meets Large-Scale Computing Tutorials 14 Sep 2018 Jeffrey A. F. Hittinger Director LLNL-PRES-758289 This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under contract DE- AC52-07NA27344. Lawrence Livermore National Security, LLC Thanks to § David Keyes, KAUST § Rob Neely, LLNL § And others… 2 LLNL-PRES-758289 Advancements in HPC have happened over three major epochs, and we’re settling on a fourth (heterogeneity) ATS-5 1.00E+19 ATS-4 Exascale 1.00E+18 Crossroads Transitions are Code 1.00E+17 Sierra essential for Transition Roadrunner 1.00E+16 Period BlueGene/L Petascale progress: between Trinity 1.00E+15 1952 and 2012 Red Storm Sequoia 1.00E+14 (Sequoia), NNSA saw Purple a factor of ~1 trillion White Cielo 1.00E+13 Terascale improvement in peak Blue Pacific After a decade or 1.00E+12 speed ASCI Red so of uncertainty, 1.00E+11 history will most 1.00E+10 Gigascale likely refer to this 1.00E+09 next era as the 1.00E+08 heterogeneous era 1.00E+07 Megascale 1.00E+06 1.00E+05 SerialSerial EraEra VectorVector EraEra DistributedDistributed Memory Era Era ManyHeterogeneous Core Era Era 1.00E+04 Kiloscale peak 1.00E+03 Jul-86 Jul-18 (flops) Jan-52 Jun-54 Dec-56 May-59 Nov-61 May-64 Oct-66 Apr-69 Sep-71 Mar-74 Aug-76 Feb-79 Aug-81 Jan-84 Dec-88 Jun-91 Nov-93 May-96 Nov-98 Apr-01 Oct-03 Mar-06 Sep-08 Feb-11 Aug-13 Jan-16 Jan-21 Jun-23 Dec-25 Each of these eras defines a new common programming model and transitions are taking longer; we are entering a fourth era 3 LLNL-PRES-758289 Pre-WWII – the theoretical foundations of modern computing are built in the U.S. and England § 1830’s – The Difference Engine — Charles Babbage / Ada Lovelace — Not built during his lifetime § 1930’s – The Turing Machine — Alan Turing (England / Princeton) — Theory of programmable algorithms § 1940’s – ENIAC — Mauchly/Eckert – Moore School of Eng. (Philadelphia) — 1st general programmable electronic computer § 1940’s – EDVAC design. — John Von Neumann (Los Alamos) — Basis for “stored program architecture” still used today Ballistics projectile calculations were the first “killer app” for computers, and were historically done by a pipeline of human “computers” 4 LLNL-PRES-758289 4 The “mainframe” era – general purpose computers designed for scientific computing 1953 § Univac 1 — First machine installed at LLNL in 1953 1954 § IBM 701 — Installed at Los Alamos also in 1953 — Williams tubes (fast memory, but unreliable) 1956 § IBM 704 — Core memory, floating point arithmetic, CRT — Commercially successful § IBM 709 1958 — Seamless porting § IBM 7090 — Transistor-based — Large speed increases § Univac LARC 1960 — Co-designed with, and for, LLNL — One of the first transistor-based machines 5 LLNL-PRES-758289 5 The “mainframe” era – continued 1961 § IBM 7030 (Stretch) • Competitor to LARC • Considered a failure at the time (only achieved 50% of performance Goals) • Introduced many concepts that went into the IBM System/360 § CDC 1604 1962 • First desiGned by Seymour Cray § CDC 3600 • 48-bit words (hiGher precision) § CDC 6600 • Considered the first real “supercomputer” (1st at CERN) 1964 • Full separation of input/output from computinG (precursor to RISC desiGns) § CDC 7600 • Hierarchical memory desiGn 1969 • Instruction pipeline (allowed for modest parallelism) • Fastest computer in world from 1969-1975 6 LLNL-PRES-758289 6 Programming in the mainframe era § Programming was often an intimate exercise with the particular computer you were targeting § Entirely serial programming § Lack of sufficient memory was often the limiting factor in scientific software design § This era saw huge innovations in hardware and software, despite computer science not yet being an established discipline — Development of compilers (e.g. Fortran) — Operating systems (e.g. CTSS) — Batch environments and time-sharing — Use of terminals for viewing output — LLNL’s “Octopus” network § U.S. stood largely alone in dominating the HPC market 7 LLNL-PRES-758289 7 The vector era 1976 § CDC Star 100 • First CDC design not led by Seymour Cray • One of the first commercial vector machines • Hard to achieve anything near peak performance • Large disparity between serial and vector code 1975 § Illiac IV (Burroughs & Univ of Illinois) • First massively parallel machine portends future trends • Moved to NASA Ames in 1975 (defense at a public university? Bah) 1978 § Cray 1 • First machine developed at Seymour Cray’s new company • Applications saw immediate gains in performance (vs CDC 7600), and could incrementally improve with additional vectorization § Cray X-MP 1983 • Designed by Steve Chen at Cray • First multi-processor supercomputer 8 LLNL-PRES-758289 8 The vector era – soon dominated by Cray in the U.S. 1985 § Cray 2 • Seymour’s 2nd design • Flourinert-cooled • Most performance gains were due to very fast memory Other players in vector computer design: 1988 § Cray Y-MP US • Successor to X-MP • Convex • First delivered with Unix-based UNICOS OS • Thinking Machines • ETA (spinoff of CDC) • IBM 1991 § Cray C90 Non-US 1994 § Cray J90 • NEC • “Mini” supercomputer – helped bridge gap into MPP era • Fujitsu • Hitachi 1995 § Cray T90 • End of an era! 9 LLNL-PRES-758289 9 Programming vector computers in that era § Machines were often delivered with no operating system—just hardware — Labs created their own software environments — Took years for the Cray compiler to catch up to the NNSA lab compilers § Vector computers are designed to work with arrays and loops — Lots of those in scientific codes! § Vectorizing was difficult, and many codes did not lend themselves to this — Loop dependencies (e.g. fib[i] = fib[i-1] + fib[i-2]) — Conditional statements — Non-constant stride through memory — I/O or system calls — Function calls Machine details matter for effective use… 10 LLNL-PRES-758289 Programming in the vector era – cont’d § Compilers became increasingly sophisticated at being able to perform some code transformations — Yet “automatic parallelization” never fully achieved its academic promises (Acquired by Intel, 200) § Vectorized portions of code could realize anywhere from 2 – 50x speedups over a serial counterpart § Vectorization (via SIMD units) eventually made its way into commodity CPUs with the Pentium III (SSE) — Also, AVX, ARM Neon, Power QPX — (Not quite the same, but close enouGh…) § Stable programming model, but — Lots of effort for potentially limited payoff — Vectorized codinG could Get uGly! Vectorization lives on in the form of SIMD units (e.g., Intel AVX) 11 LLNL-PRES-758289 11 Soon…mainframes (vector or not) faced the “attack of the killer micros” § Late 1980s; heavy influence by the PC/workstation market — Reduced instruction set computing (RISC) processors dominated price-performance — Mainframes became “dinosaurs running COBOL” § Revolution driven largely by cost and scaling — Cost: computing went mainstream — Scaling: commodity processors could be interconnected — Programming: rewrite from vector code to commodity processors Penetrate lower end Penetrate high end market through (HPC) market low-cost commodity through parallel CMOS processors scaling With industry focusing its investment on commodity processors, the writing was on the wall for the labs: leverage or fall behind the curve… 12 LLNL-PRES-758289 Commodity hardware components thus defines the distributed memory era § Workstations (e.g. Silicon Graphics) were doing the work of mainframes for 50x cheaper! — SGI acquires Cray Research § The relative ease of building distributed memory systems launched a wave of innovation and new supercomputer companies in the 90’s — IBM, Cray [Research, Computer, Inc], SGI, Intel, nCube, MasPar, Alliant, Multiflow, Kendall Square, BBN, DEC, Sequent, Convex, Encore, HP, Meiko, Supertek, Floating Point Systems, Tera, MIT J-machine, … § Fast commodity Ethernet and Linux eventually lead to the “Beowulf” explosion — Anyone can build a supercomputer now! — Great boon to universities and small companies — By the 2000’s, the list of dead or dying U.S. Supercomputing companies exceeded those still alive 13 LLNL-PRES-758289 13 The distributed computing or MPP era § ASCI program launches in ~1995 — Replace underground nuclear testing with science-based stockpile stewardship (simulation and experimental facilities) § Meanwhile, the Japanese “stay the course” with vectors — Gov’t funding pumped up their market — Some concern that US had forever lost it’s edge when the Japanese Earth Simulator was deployed in 2002. — Many here deemed it a Pyrrhic victory • It had a HUGE footprint and power use § BlueGene/L (@ LLNL) retakes the crown in 2004 for 3.5 straight years, putting a nail in the coffin of the Japanese dominance § DOE Leadership Computing Facilities (ANL/ORNL) launched in ~2004 to help reestablish US lead in open scientific computing — Currently housing Summit (#1 @ ORNL), Mira (#17 @ ANL) — Argonne scheduled to receive first exascale computer in 2021-22 (A21) 14 LLNL-PRES-758289 14 Programming in the Distributed Computing Era § Going on 20+ years of stability in programming model — PVM, and then ultimately MPI, provide performance portability — Unix becomes the de-facto standard for HPC — Object-oriented design helps balance complexity and performance — Programming Languages and compilers continue to improve — Emergence of grid and distance computing § Mission-delivery and scientific advancement has advanced rapidly,
Details
-
File Typepdf
-
Upload Time-
-
Content LanguagesEnglish
-
Upload UserAnonymous/Not logged-in
-
File Pages47 Page
-
File Size-